SynthScope / LEADERBOARD.md
SMSD75's picture
Update SynthScope Space from the GitHub repository
cd24b4a verified
|
Raw History Blame Contribute Delete
2.89 kB

SynthScope Leaderboard (v1)

Auto-generated by benchmark/leaderboard.py β€” do not edit by hand. See benchmark/README.md for the protocol and how to submit.

This is the standardized, versioned table: a few tracks with a fixed protocol that anyone can submit to. The web page's Leaderboard tab is broader (every paper benchmark, 30+ groups) and the repo supports many more datasets, which download and organize themselves: see docs/BENCHMARKS.md.

Metrics are percentages. mAP = mean AP over generators. Each track is ranked by its primary metric (in bold).

ForenSynths β€” image (ProGAN-trained, cross-generator)

Rank Method Eval Acc mAP AP AUROC Verified Notes
1 FatFormer official-weights 98.39 99.70 β€” β€” βœ“ ForenSynths-8 subset; official released weights
2 CO-SPY (official weights) official-weights 95.12 98.71 β€” β€” βœ“ ForenSynths-8; vendored official architecture
3 UnivFD / CLIP frozen-probe β€” 97.20 β€” 97.00 βœ“ ForenSynths; frozen CLIP ViT-L/14 probe (ProGAN)
4 CO-SPY (from scratch) from-scratch β€” 96.98 β€” 96.79 βœ“ ForenSynths; SynthScope re-implementation trained on ProGAN
5 CNNDetection from-scratch β€” 96.80 β€” 96.60 βœ“ ForenSynths; ResNet-50 fine-tune (ProGAN)
6 DINOv2 probe frozen-probe β€” 95.03 β€” 94.86 βœ“ ForenSynths; frozen DINOv2 ViT-L/14 probe
7 SAFE from-scratch β€” 94.13 β€” 93.68 βœ“ ForenSynths; trained from scratch
8 AIDE official-weights β€” 92.08 β€” β€” βœ“ ForenSynths-8; mAP reproduces, acc@0.5 miscalibrated (single-view vs official 5-view DCT)
9 NPR official-weights 91.83 87.09 β€” β€” βœ“ ForenSynths-8; official ProGAN weights

GenImage SD-v1.4 β€” image

Rank Method Eval Acc mAP AP AUROC Verified Notes
1 VIB-Net from-scratch β€” 99.33 β€” 99.33 βœ“ GenImage SD-v1.4 single real/fake pool (mAP == AUROC by construction)
2 SPAI (official weights) official-weights β€” β€” 94.01 98.45 βœ“ GenImage SD-v1.4; vendored official architecture, fixed-res 2x2x224
3 SPAI (from scratch) from-scratch β€” β€” β€” 97.66 βœ“ DMID Latent-Diffusion vs COCO; SynthScope fixed-resolution variant

GenVideo-Val β€” video

Rank Method Eval Acc mAP AP AUROC Verified Notes
1 ReStraV frozen-probe β€” β€” 96.67 97.06 βœ“ Frozen DINOv2 trajectory geometry + MLP
2 D3 training-free β€” β€” 95.60 93.60 βœ“ Training-free frozen CLIP 2nd-order trajectory